Debugging in C++ codes

[Alex Jiang]

Hi all,

I am trying to explore the detailed information in C++ codes for post training quantization. But when I started step-by-step debugging, I can only reach TocoConvert

in tensorflow/lite/python/wrap_toco.py:32 and can not step in to this API. I was wondering if anyone has any suggestions? 

Also, if this can not be done, is there any solution to convert a model (.ckpt) to a tflite model using C++?

Thank you!

[Jaesung Chung]

Hi,

For the .ckpt conversion, we provide the saved model conversion API through tf.lite.TFLiteConverter.from_saved_model (link). Please load the saved model with checkpoints by using the TensorFlow APIs and then, convert the loaded saved model to the above converter API.

For the detailed quantization code, you can check out from the MLIR code (https://github.com/tensorflow/tensorflow/blob/master/tensorflow/compiler/mlir/lite/quantization/lite/quantize_model.h#L32)

Best regards,

Jaesung

[Alex Jiang]

Hi Jaesung, 

Thank you for your information! 

I have converted the model using  API through tf.lite.TFLiteConverter.from_saved_model. I just wonder if there is any solution to convert by C++ or call functions in C++ codes, since I really curious about the dataflow.

Best,

Alex

[Jaesung Chung]

Hi Alex,

Good to hear that the conversion has succeeded already.

There are no official APIs for the TFLite converter layer in the C++ level. However, if you would like to know about the conversion process details, you can check out the TensorFlow Lite MLIR converter here.

Best regards,

Jaesung

[Alex Jiang]

Hi Jaesung,

I will do that. Thank you so much!

Best,

Alex